Humans are still significantly better at predicting the weather than AI

[Polycount]

In context: Artificial intelligence is present in almost every aspect of our daily lives. It's used in search engines, email clients, media platforms, and even grocery chains worldwide. It arguably makes the world a more efficient place, but not without costs; job loss being one of them. However, there is one industry that still struggles to fully automate: weather forecasting.

Despite all the progress that has been made in the realm of technology, AI, and machine learning, weather companies have not yet handed over the keys to their computer counterparts. While AI models, with the help of weather satellites like GOES-16 and 17, are capable of predicting minor weather changes and patterns with surprising accuracy, their efficiency takes a steep dive when major climate disruptions occur.

These major weather shifts, including natural disasters, often have warning signs that are too subtle and small in scale for most models to detect. Even when they are detected, models can't always draw correlations between early indicators of, say, a waterspout and its likely arrival. This, Wired reports, usually requires the practiced eye of an industry veteran. The outlet claims that, based on over two decades worth of weather prediction information (gathered by the NOAA Weather Prediction Service), humans outperform two of the most popular national weather prediction models; the Global Forecast System and the North American Mesoscale Forecast System.

By how much? Between 20 and 40 percent.

That is no small gap, and it could easily mean the difference between life and death for those in the path of a vicious tornado, waterspout, or rapid-onset blizzard. Having access to accurate weather information as early as possible is what makes evacuation (where necessary) or shelter-in-place recommendations viable.

Seasoned weather veterans can look at the smallest details, like subtle shifts in atmospheric pressure, wind speed, or "available moisture," and draw higher-quality conclusions than their computerized counterparts. This is because, in many cases, weather prediction models don't weigh these measures as highly.

And frankly, that's no surprise. Computers are smart and getting smarter by the day, but they still lack something humans have always possessed: the ability to evaluate situations in a broader context. Bringing machines anywhere near parity with human memory and context awareness would require an immense amount of processing power that just isn't available widely at the moment. Only a handful of computers that might be capable of this are in development right now -- at least, in the US -- and weather groups aren't the only ones that want to use them.

In short: humans are actively being replaced across dozens of industries at an alarming rate, but for now, your local weatherman (or at least the group feeding him his data) is safe.

Permalink to story.

https://www.techspot.com/news/92936-humans-significantly-better-predicting-weather-than-ai.html

 

[Mjsun]

I have a bone to pick with the writer and editor:
The article title claims that AI is no better than a human for predicting the weather, but then the article only mentions two weather models (GFS and NAM) that as far as I can tell feature no neural nets/GANs/deep learning. As such, the two models are not AI. They are incredibly complicated but classical finite element models. If the author is aware of info on the models that shows the use of AI, then please add that and some additional primer material into the article. Otherwise, I think the article title is misleading.

 

[tellmewhy]

In the passage of time in the atmosphere different conditions occur and in each condition a different subset of characteristics and ONLY this subset plays a role in the form of weather. Thus, when they train the network in all the characteristics (temperature, humidity, pressure, wind, time, sunlight, etc.) they train it in data with a lot of noise, so the neural network due to noise can not distinguish the patterns within the data set.

They therefore need a neural network to determine which characteristic set is dominant in each state of the atmosphere to play a role in the weather and 10 or more neural networks for each dominant combined feature set.

Those who usually use neural networks are a bit lazy, they think that they do not need to think about just using the neural network to brute force the solve of the problem. As we can see, they still have to think

 

[Plutoisaplanet]

I just checked the source article's sources, and the 20-40% number is actually 7.5 years old: https://journals.ametsoc.org/view/journals/wefo/29/3/waf-d-13-00066_1.xml

In addition, the prediction of waterspouts being reported is based on data 5 years old at this point: https://journals.ametsoc.org/view/journals/wefo/33/2/waf-d-17-0100_1.xml

The only thing that the source Wired article reported that was new (as far as I can tell) is all the qualitative stuff like quotes from meteorologists, and exploring the shortcomings in detail in one place.

 

[Lounds]

Red sky in morning Sheppard's warning.
Red sky at night Sheppard's delight.

 

[Godel]

I notice that the two systems mentioned are both American. In other articles I've read that the European models are often more accurate in predicting hurricane tracks and the like.

 

[Mr Majestyk]

Come to Australia and I'll show you how much humans suck at weather forecasting.